Gideon
douatwork/gideon
Integração cross-agent com sincronização bidirecional de conhecimento
Ver no GitHubHarness leve que integra assistentes de código (Claude Code, Pi Agent) como assistentes pessoais. Oferece uma skill única que cuida de scaffolding, bootstrap e sincronização bidirecional de workflows entre agentes e repositório central.
Instalação
npx convez-ai douatwork/gideon1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Orquestrar múltiplos agentes de código com um ponto de entrada unificado
- Sincronizar workflows e skills entre agentes automaticamente
- Scaffolding e bootstrap de novos ambientes de runtime de agentes
Veredito

Solução bem arquitetada para operadores que precisam gerenciar múltiplos agentes de IA. O design idempotente e o foco em entrada de usuário descoberta aumentam a usabilidade, mas requer familiaridade com estrutura de agentes e git.
README
A tradução para o português deste README ainda está na fila.
Gideon — single cross-agent skill
This folder is one skill that does everything operational in Gideon from a single invocation:
- Scaffolds the
core/ stack/ taste/ workflows/folders (plusAGENTS.md,.sync/, and a git repo) when Gideon doesn't exist yet. - Bootstraps a new agent runtime — writes the user-level root instruction file that points at Gideon and registers a per-agent state file.
- Syncs bidirectionally — contributes what the agent knows back into Gideon, then derives native skills from Gideon's workflow files and records the synced commit.
It collapses Gideon's three otherwise-separate concerns (workflows/agents-init.md,
workflows/agent-sync.md, and the previously-manual folder setup) into one
idempotent, mode-detecting skill. Run it any time; it does only what's needed.
Why a single skill
Gideon's repo keeps agents-init.md and agent-sync.md as system instructions
that each runtime reads ad hoc and never turns into a skill. That works, but it
leaves the operator-facing entry points scattered and requires the user to know
which document to point an agent at. Folding them into one installable skill
gives every runtime the same first-class verb — invoke gideon and it figures
out whether to scaffold, bootstrap, or sync. The cost is one derived artifact per
runtime; the payoff is a uniform, discoverable entry point.
Installing it per runtime
SKILL.md is written runtime-agnostically (name + description frontmatter,
instruction-only body). Drop it into each runtime's skills directory:
| Runtime | Where it goes | Notes |
|---|---|---|
| Claude Code | ~/.claude/skills/gideon/SKILL.md | Frontmatter name/description is enough. |
| pi | runtime skills dir, gideon/SKILL.md | Same SKILL.md frontmatter contract. |
| Codex | gideon/SKILL.md + gideon/agents/openai.yaml | Codex needs the extra YAML; mirror name/description into it. |
| other | the runtime's skills/commands dir | Any runtime that loads a markdown skill with a name + description trigger works as-is. |
Bootstrapping is a chicken-and-egg only once: copy SKILL.md into the runtime,
then invoke it (e.g. "run the gideon skill"). From there it writes the root
instruction file, so later sessions discover Gideon on their own and you just say
"sync with Gideon" when things change.
Relationship to the canonical workflow files
This skill is a faithful superset of public/workflows/agents-init.md and
public/workflows/agent-sync.md — same git-safety discipline, same bidirectional
phase order, same state-file mechanics — plus the scaffold step those files
assume has already happened. If the upstream sync/init contract changes, update
SKILL.md to match; it is the derived artifact and Gideon's workflow files remain
the source of truth.